README.md

---
title: ACE-Step WebGPU
emoji: 🎵
colorFrom: purple
colorTo: pink
sdk: static
pinned: false
license: apache-2.0
short_description: Text-to-music in your browser via WebGPU.
---

# ACE-Step WebGPU

Describe any song. AI writes & produces it — right in your browser.

The pipeline (5 Hz Qwen3 LM → FSQ → DiT decoder → Oobleck VAE) runs end-to-end
via [onnxruntime-web](https://onnxruntime.ai/) with the WebGPU execution
provider. Two Web Workers keep the LM and diffusion+VAE graphs in separate
WASM heaps so neither hits the 4 GB single-heap limit.

## Models

- DiT decoder (2B, fp16) and Oobleck VAE (fp16) from
  [shreyask/ACE-Step-v1.5-ONNX](https://huggingface.co/shreyask/ACE-Step-v1.5-ONNX)
- 5 Hz LM (0.6B, 4-bit MatMulNBits) from
  [ACE-Step/acestep-5Hz-lm-0.6B](https://huggingface.co/ACE-Step/acestep-5Hz-lm-0.6B)
- Text encoder: [Qwen/Qwen3-Embedding-0.6B](https://huggingface.co/Qwen/Qwen3-Embedding-0.6B)

Weights are fetched on demand and cached in the browser's Cache Storage after
the first load (~2 GB total).

## Requirements

- WebGPU-capable browser: Chrome/Edge 113+, Safari 26+ desktop
- ~4 GB free RAM recommended

## Source

The `_source/` directory in this Space's Files tab contains the full Vite/React
project (`src/`, `public/`, configs). Build it locally with:

```bash
cd _source
npm install
npm run build
```

Upstream: [ACE-Step/Ace-Step1.5](https://huggingface.co/ACE-Step/Ace-Step1.5)
(Apache 2.0).